40 results about "Visual task" patented technology

The application discloses a mirror highlight detection and removal method based on a double-flow convolutional neural network, which calculates the gradient of a pixel point in an x direction and a y direction of an input original image with highlights, extracts a first highlight feature mapping, subtracts the first highlight feature mapping after being processed by a first convolution block attention module CBAM from a preprocessed image, and outputs a first-stage highlight-free image; the first highlight feature mapping is progressively down-sampled and reduced in size, each highlight feature mapping after being down-sampled and reduced in size is processed by a convolution block attention module CBAM at each stage, and is subtracted from a highlight-free image output by a previous stage, and finally a rough highlight-free image is obtained. Then, highlight extraction and refinement are performed on the rough highlight-free image by a highlight extraction module to obtain a final highlight-free image. The application can effectively solve the image information degradation problem caused by the mirror highlight, thereby reducing the interference of the highlight on visual tasks such as target detection.

This invention discloses a self-supervised depth completion method, system, and terminal based on 3D feature extraction and fusion. The method includes: acquiring a discrete depth map and a color map; preprocessing the discrete depth map to obtain a target discrete depth map; downsampling the color map to obtain a target color feature map; performing a preset number of downsampling and feature extraction operations on the target discrete depth map and the color map to obtain a first color feature map and a first target discrete depth feature map; performing channel concatenation and upsampling operations to obtain a fused image feature map; inputting the fused image feature map and the target discrete depth map into a cross-attention feature fusion module to output the fused feature map; and performing channel concatenation and upsampling on the target color feature map and the fused feature map to obtain a completed depth map. This invention can obtain a completed depth map after information completion, thereby enabling accurate processing of subsequent computer vision tasks.

This invention relates to the field of artificial intelligence technology and discloses a linear attention mechanism method, device, and electronic device for visual tasks. The method includes: performing norm decomposition and direction decomposition on the input query vector and key vector respectively to obtain the magnitude and unit direction vector corresponding to the query vector and key vector; designing a query norm-driven dynamic entropy reduction control mechanism and a cosine suppression nonnegativity preservation mechanism to obtain the query kernel function feature vector and the key kernel function feature vector; and finally calculating the linear attention result based on the query kernel function feature vector and the key kernel function feature vector. This invention significantly improves the model's ability to process information of different intensities while maintaining linear computational complexity, enhances the model's performance in complex tasks, and thus achieves a balance between efficiency and expressive power, effectively solving the problem of the lack of correlation between norm and entropy value in traditional linear attention mechanisms.

This invention provides a multi-task AI intelligent agent cost estimation system and method adapted to hydropower engineering cost estimation scenarios, relating to the field of hydropower engineering cost estimation technology. It utilizes a visual task interaction module to input basic hydropower engineering information and multi-task instructions. A multi-task AI intelligent agent decouples the input basic hydropower engineering information and multi-task instructions to obtain a task list and task priorities. Based on the tasks, it calls the corresponding data resource sub-libraries, and according to the corresponding calculation logic and priorities, combines the data resource sub-libraries to complete multi-task calculations, outputting the calculation results for each task. This invention solves the problems of information silos, low efficiency, and high error rates in existing cost estimation systems and is applicable to hydropower engineering cost estimation.

The disclosure provides a self-supervised learning method and a self-supervised learning device, and relates to the field of computers. The disclosure can effectively monitor the augmented picture samples with semantic deviation through the distance measurement information of the feature of each augmented picture of each original picture to the feature mean of all augmented pictures of the original picture, and the training noise caused by the augmented picture samples with semantic deviation can be effectively inhibited by reducing the corresponding weight, thereby balancing the variance and deviation of the data augmentation distribution, and improving the performance of the learning model in downstream computer vision tasks.

The application belongs to the field of image processing, and particularly relates to a block convolution calculation method, system and device for a visual task, aiming to solve the problem that the calculation resource occupation of the existing graphic processing method is too large, resulting in that the neural network model is difficult to apply in an embedded device with insufficient size or performance. The application comprises the following steps: acquiring a to-be-processed feature map by an image acquisition device, and marking the feature map as a 0th layer feature map; marking the layer number of the current feature map as i, at this time i=0; equally dividing the current feature map into a plurality of original blocks with a preset size; performing edge zero padding based on the original blocks to obtain zero padding blocks; and performing calculation on the zero padding blocks one by one through a convolution layer to obtain an (i+1)th layer feature map. According to the application, the calculation of each layer of convolution is split into several independent block convolution calculations, and compared with an ordinary convolution model, the required memory is smaller in the case of the same model accuracy.

The application discloses a kind of target detection methods for unmanned aerial vehicle aerial photograph, belong to image data processing technical field, its steps are: the input unmanned aerial vehicle aerial photograph is preprocessed;The image after pre-processing is input into main network and carries out multi-stage feature extraction, and the main network is five-stage progressive feature extraction architecture, and its core component is feature extraction residual module CSP-CGLA;The feature extracted by main network is input into neck network and is fused from top to bottom, to obtain shallow layer fusion feature;Shallow layer fusion feature is progressively down-sampled and fused, and enhanced middle layer fusion feature and deep layer fusion feature are sequentially generated;Output multi-scale enhanced feature is output, for downstream visual task, significantly improve visual perception accuracy and efficiency on unmanned aerial vehicle platform.

This invention, a multimodal underwater scene generation method based on a diffusion model (UMDM-USG), belongs to the field of computer vision and image generation technology. Addressing the high cost of underwater scene data acquisition, this invention first encodes the features of the input text and multimodal data. Second, it randomly divides the modalities into generative and conditional modalities through a role assignment module. Then, it constructs a generative conditional alignment attention module in the multimodal alignment module to achieve bidirectional interaction and alignment between the generative and conditional modalities. Finally, it outputs the results of each modality through a decoder. Furthermore, a representation alignment regularization strategy is introduced, utilizing the visual prior of a pre-trained self-supervised model as a supervision signal to guide the generation distribution towards a realistic scene, significantly improving geometric fidelity. Experimental results show that this invention outperforms existing methods in terms of generation quality, semantic consistency, and multimodal collaboration, and can provide reliable data support for downstream vision tasks.

The application discloses a pre-cataract surgery visual function expectation evaluation method, collects preoperative basic examination data, subjective expectation expression data and a preset visual task scene library, and screens a candidate visual task scene set; a scene acceptance result set is obtained through patient acceptance interaction discrimination; subjective expectation expression data are combined and analyzed to generate a visual function requirement parameter set; a visual task tolerance threshold is further quantified and structured and coded into an expectation feature vector, and a preoperative visual function expectation evaluation result is output, which is used for preoperative lens matching, preoperative communication prompting or postoperative satisfaction risk prediction.

A computer-implemented method and system refer to an image encoder that receives a digital image as input. The image encoder generates a weighting map using a preceding feature map. The preceding feature map is generated using pixels of the digital image. The weighting map is generated based on Lie data associated with the digital image. A homographic transformation is interpolated between two planar projections of the digital image using at least the weighting map and a homography matrix. The homography matrix provides a mapping between the two planar projections of the digital image. Homographically transformed kernels are generated by applying the homographic transformation to convolution kernels.The homographically transformed kernels are applied to the preceding feature map to perform folding on different planar areas appearing in the digital image and to generate a new feature map that is used for a computer vision task involving three-dimensional (3D) perception.

This invention discloses a single-image robot visual image detail enhancement algorithm inspired by exponential-trigonometric functions, comprising the following steps: Step 1, obtaining initial residual features from the original image; Step 2, designing a loss function to describe pixel value loss and evaluate whether the residual image patch matching reaches the optimal level; Step 3, initializing image candidate patches, setting search boundaries and the initial position of each image candidate patch, and comparing each image candidate patch with the designed loss function to obtain the initial optimal image patch and the initial optimal fitness; Step 4, using an optimization algorithm to iteratively search the original image using the initial residual features and perform residual updates to obtain the final optimal image patch; Step 5, superimposing the optimal image patch onto the original image to obtain the final detail-enhanced image. This invention can significantly improve the texture clarity and edge discernibility of robot-acquired images, directly improving the accuracy and stability of downstream visual tasks.

Systems and methods for privacy-preserving optics are described. An embodiments includes a method of preserving-privacy on captured images while performing a computer vision task that includes generating an optimal set of parameters to parameterize an encoding optical element to produce optical distortions such that acquired images by the camera are distorted, where the optimal set of parameters is learned via end-to-end learning that jointly optimizes from a camera optics model to a computational process that performs a computer vision task on the distorted images acquired by the camera, where the distorted images visually obscure a privacy attribute of people to protect their privacy but still preserve features to perform the computer vision task, acquiring several distorted images, and performing a computer vision task directly on the distorted images where the distortions generated by the camera are optimal and allow obtaining high performance on the computer vision task.

The application discloses a kind of multi-modal image's synergic registration fusion method, system, electronic equipment and storage medium, the method includes: S1 obtains the multi-modal image group of target area, the multi-modal image group is divided into full-color image and multispectral image image;S2 the full-color image and multispectral image image are preprocessed, obtain the full-color image and multispectral image image after preprocessing;S3 the full-color image and multispectral image image after preprocessing are input into synergic registration fusion network model, obtain target fusion image.The synergic registration fusion network model constructed by the application includes image registration network, semantic extraction network and image fusion network, can effectively eliminate the artifact caused by parallax between multi-modal image, and introduce semantic feature in fusion image, improve the quality of fusion image, meet the demand of subsequent advanced visual task.

This invention relates to the field of deep learning technology and discloses a method for training a human visual pedestal model, a human body weight recognition method, a device, an apparatus, and a storage medium. This invention constructs a self-supervised learning human visual pedestal model, which can be trained without labeling training human images. Only a massive amount of training human images need to be input into the human visual pedestal model, allowing it to undergo self-supervised training based on these images. Furthermore, after the human visual pedestal model is trained, fine-tuning can be performed by adjusting and freezing the control factors in the encoder, making it applicable to downstream human-centered vision tasks. Compared to existing technologies that train human visual pedestal models based on small-scale labeled data or labeled data from specific scenarios, the human visual pedestal model trained by this invention exhibits stronger generalization capabilities.

The invention discloses an edge visual task dynamic unloading method based on multi-dimensional perception, relates to the technical field of visual tasks, and is used for solving the problems of unstable result output and out-of-control triggering time sequence caused by improper selection of local processing and edge unloading in a visual task execution process. Boundary information of a monitoring area is obtained, a warning area is constructed, a trigger welt index is generated in combination with a target staying feature and a boundary approaching feature so as to identify a warning edge stage, processing time information of a video analysis task is recorded in the warning edge stage, and processing sequence features are evaluated. A target detection frame is extracted, a target shielding ratio is calculated, processing sequence features and a shielding change trend are fused to construct a task identification state, dynamic switching is performed between a local processing path and an edge processing path based on the task identification state, and cooperative balance of identification precision requirements and result stability is realized. And stable output, consistent sequence and controllable triggering time sequence of the identification result in the key stage are ensured.

The application discloses a kind of high-efficiency differentiable neural architecture search methods based on topological structure optimization, comprising the following steps: according to computer vision task classification form, determine model structure and search network, form the computational graph of searchable structure; The architecture parameters of computational graph are determined by structure parameters and network parameters;Iterative optimization is carried out to the topological structure of computational graph, redundant structure is eliminated, and structure parameters are optimized;Carry out convolution operation;Optimize network parameters by convolution kernel reparameterization;And by convolution kernel standardization, decouple network parameters and structure parameters;Gradient descent algorithm is obtained after the optimized computational graph runs, and the optimal architecture parameter is determined;Train model structure, complete computer vision task classification.This method simplifies the network topological structure of differentiable neural architecture search, and sets the optimization algorithm of simplified structure targetedly;Help to complete network light weight and neural architecture search in complex visual task application.

The application provides a multi-vision task accelerator and a control method for multi-vision task processing, wherein the vision task refers to a task of processing input images by using a convolutional neural network, each image convolutional neural network comprises at least one convolutional layer, the accelerator comprises: a calculation array for performing convolution operation of the convolutional layer; and a controller for controlling the calculation array to perform convolution operation corresponding to one vision task or dividing the calculation array into at least two areas to simultaneously perform convolution operation corresponding to at least two vision tasks in the multiple vision tasks when there are idle calculation resources in the calculation array performing convolution operation corresponding to a single vision task in response to an acceleration calculation request of one or more vision tasks.

3D object detection is a computer vision task that generally detects (e.g. classifies and localizes) objects in 3D space from the 2D images or videos that capture the objects. Current techniques used for 3D object detection rely on machine learning processes that learn to detect 3D objects from existing images annotated with high-quality 3D information including depth information generally obtained using lidar technology. However, due to lidar's limited measurable range, current machine learning solutions to 3D object detection do not support detection of 3D objects beyond the lidar range, which is needed for numerous applications, including autonomous driving applications where existing close or midrange 3D object detection does not always meet the safety-critical requirement of autonomous driving. The present disclosure provides for 3D object detection using a technique that supports long-range detection (i.e. detection beyond the lidar range).

The present disclosure relates to systems and methods for active testing based on visual transformers for label efficient evaluation of visual tasks. One method includes dividing an input image into a plurality of patches using a visual transformer, where each patch corresponds to a different region of the input image. The input image is defined using model outputs of a vision model. The method further includes defining a plurality of position embeddings using the visual transformer, the plurality of position embeddings including a position embedding for each of the plurality of patches and for the input image as a whole, labeling identified regions of the original image based on an estimated loss map to define a labeled image; and outputting a test performance qualifier indicative of an expected performance of the vision model when the vision model is part of a vision system. The test performance qualifier is computed using a weighted analysis based on an image loss level and a region loss level for each patch the image is provided to be labeled.

The application provides a kind of graph data analysis system and method for computer vision, by obtaining the input image to be analyzed, and extracting initial visual feature map from input image;According to the visual semantic feature similarity and relative geometric distance between the feature voxel entities corresponding to each spatial position in initial visual feature map, the initial relationship graph of the potential semantic correlation of input image is constructed;The response characteristics of the intermediate layer of the deep convolution network are recalibrated based on the initial relationship graph, and the enhanced visual feature representation is generated;The context feature modulation of enhanced visual feature representation is carried out based on the optimized relationship graph determined by enhanced visual feature representation, and the target visual feature map is generated;Based on the visual-topology fusion structured representation determined by target visual feature map and optimized relationship graph, the predetermined computer vision task is executed.The technical scheme provided by the application can analyze the visual features of the input image to be analyzed under the influence of inconsistent image semantic context.

This invention discloses a monitoring image enhancement system and method based on a large visual model under complex meteorological conditions, belonging to the technical field of image enhancement. The system includes a data preprocessing module, a degradation phenomenon recognition model fine-tuning module, a degradation suppression processing module, and a downstream task adaptation module. The data preprocessing module performs spatiotemporal standardization on the monitoring video, unifying the frame rate and resolution, and extracting segments with fixed frame lengths. The degradation phenomenon recognition model fine-tuning module fine-tunes the basic large visual model based on multi-type meteorological degradation datasets, achieving accurate identification of degradation types, degradation regions, and color temperature parameters. The degradation suppression processing module adopts a layered processing strategy to complete targeted removal of meteorological degradation, enhancement of high-frequency residual details, and color and contrast correction. The downstream task adaptation module connects the enhanced image to a downstream visual task model. This invention can simultaneously suppress multiple degradation interferences such as rain, snow, fog, haze, and low light, effectively restoring image details and colors.

This invention relates to the field of computer vision architecture design technology, specifically disclosing a visual feature modeling method based on global memory perception and local feature decoupling. The method includes: constructing a dual-branch parallel architecture to structurally decouple global information representation from local feature extraction functions; generating fine-grained local features in the local branch using large-kernel depthwise convolution and lightweight channel hybrid operations; performing feature compression in the global branch using a "feature compression-blocking-reassembly-channel restoration" strategy; modeling long-range context information for the block features using a lightweight global memory perception module to perform retrieval-update operations; and performing pixel-level modulation of the features in the local branch through a learnable gating mechanism, where the modulation process guides rather than overwrites the local features. This invention significantly improves the accuracy and efficiency of visual tasks, significantly reduces computational overhead and feature distortion problems, and provides more efficient and accurate feature modeling support for tasks such as image classification and object detection.

The application provides an OpenVX framework system for NPU acceleration, comprising an application layer, an OpenVX framework layer, an NPU runtime layer and an NPU hardware layer, wherein the OpenVX framework layer comprises an NPU-aware graph optimizer. The application designs an OpenVX framework for completing a visual task on an NPU acceleration chip, can optimize a computation graph of the visual task constructed by a standard OpenVX API according to hardware characteristics of the NPU, and designs memory management of the NPU hardware layer, so that the visual task can be accelerated and completed on the NPU, and the computation efficiency of the visual task is improved.

The application belongs to the field of real scene three-dimensional reconstruction, and provides a house three-dimensional model reconstruction method and equipment based on missing point cloud data, and a medium. The method fuses an image generated point cloud method, a neural network method, and a skeleton line extraction method, and provides a new solution idea for processing of missing point cloud data; video or panoramic image data is used to generate point cloud data based on an SFM principle; a region growing algorithm is used to sample and segment the point cloud; a neural network based on PointNet is constructed, and a cross entropy loss function is used to realize point cloud missingness judgment; a high confidence point cloud mapping completion method depending on a sampling point as a Truth Point is used to realize a full-process real scene three-dimensional reconstruction, and images are fused into a three-dimensional scene according to strict geometric relations. The application has wide use value in automatic driving, indoor navigation, augmented reality (AR), robots and other related visual task scenes based on the field of real scene three-dimensional reconstruction.

This invention provides a method and system for performance measurement under isolated task and environmental conditions, belonging to the field of human factors engineering and performance evaluation technology. The system includes: an environmental simulation and control module, a team task simulation module with an integrated visual task editor, a multimodal data synchronous acquisition module, a performance calculation engine, a real-time monitoring and intervention module, and a data playback and analysis module. This invention deeply couples customizable complex team tasks into a high-fidelity simulated isolated and confined environment, and synchronously collects multi-dimensional process data such as operational behavior, team interaction, physiological signals, and environmental states. This achieves comprehensive, refined, and multi-dimensional quantitative measurement and evaluation of team and individual operational effectiveness under long-distance voyages and isolated and confined environments, overcoming the shortcomings of existing technologies that only focus on task results and ignore process performance and environmental coupling.

A computerized behavioral measurement system, method, and computer-readable medium are disclosed. The system includes a display, one or more input devices, at least one processor, and memory storing instructions that cause the system to present a sequence of visual stimuli comprising photographs of human subjects. For each stimulus, the system enforces at least one of a presentation interval and a response window, receives a participant response during the response window, and measures an interaction metric comprising a dwell-time metric for the stimulus and / or a defined region-of-interest. The system generates and stores a machine-readable output record comprising category-segmented response metrics and may normalize such metrics to produce participant-specific normalized metrics. In certain embodiments, a second task module presents lexical stimuli and generates a unified output record including visual-task metrics and lexical error rate.

This invention belongs to the field of image processing technology and relates to a method for correcting focal length errors in zoom cameras based on image processing. The method includes: acquiring multi-level calibration board images of the zoom camera; calculating the neighborhood gradient direction entropy pixel-by-pixel based on the RGB three channels to generate an entropy-weighted multi-channel gradient field fusion image; constructing a Gaussian scale space pyramid; obtaining the initial coordinates and feature scales of candidate calibration points through multi-scale Harris-Laplace response and scale space non-maximum suppression; building a local gray-level topological model related to illumination affine distortion and scale geometric constraints; solving for the sub-pixel coordinates of the calibration points and calculating the location confidence; iteratively optimizing the calibration point coordinates and camera intrinsic parameters; and establishing a mapping model between the nominal focal length and the actual focal length after convergence to complete the focal length error correction. This invention can accurately correct the focal length error of zoom cameras and improve the accuracy of calibration and vision tasks.

This invention relates to a device and method for improving SSVEP-BCI performance based on transcranial random noise stimulation, belonging to the field of brain-computer interface technology in bioengineering. This invention achieves dynamic optimization of tRNS parameters based on real-time SSVEP signal feedback through closed-loop integration of a tRNS visual modulation module, an SSVEP signal processing module, and an adaptive feedback module. Compared with traditional SSVEP-BCI systems, this invention, through the aforementioned closed-loop adaptive tRNS modulation, improves SSVEP SNR by 25%-40% in low-contrast scenes, maintains a stable decoding accuracy at a relatively high level of 85%-90%, and reduces the SSVEP amplitude attenuation rate from 35% to below 10% in long-term visual tasks, significantly improving system robustness and user comfort.